A method for generating long text summaries based on graph neural networks

Abstract

This invention relates to the field of text summarization technology, specifically to a long text summarization generation method based on graph neural networks. The method involves preprocessing the source text to obtain word embeddings for each sentence, which are then input into a pre-trained BERT algorithm. The output of the BERT algorithm is then fed into a dilated gated convolutional network (DGCN). The outputs of the BERT and DGCNs are aggregated in a residual structure to prevent gradient vanishing. Multi-level semantic similarity maps and multi-level natural relation maps are then constructed and passed through graph attention layers. After aggregation, sentence labels are predicted using activation functions and fully connected layers, and a summarization loss function is constructed. Finally, a model loss function is calculated based on the summarization loss function and the topic model loss function, enabling the extracted text summary to more comprehensively retain the important information of the source text.

Description

Technical Field

[0001] This invention relates to the field of text summarization technology, and specifically to a method for generating long text summaries based on graph neural networks. Background Technology

[0002] In recent years, the number of research papers published in various conferences and journals has exploded. This vast amount of academic articles is a valuable source of information and knowledge. However, with the rapid iteration of science and technology, beginners or newcomers to the field often find it difficult to select suitable articles to read, and even when they do find appropriate scientific papers, it is difficult to quickly grasp the overall information. Therefore, how to effectively summarize the massive amount of scientific papers has become a key issue.

[0003] Furthermore, a large number of articles also involve fabrication and plagiarism. The PubMed database contains as many as 34 million papers, and it is estimated that at least 340,000 of them may have problems. Moreover, some plagiarized papers modify the wording but retain the core content, still falling under the category of plagiarism. In such cases, a full-text comparison may not detect the plagiarism. Therefore, generating a long-text summary to summarize the entire paper, and then comparing it with the core summary, will not only significantly reduce computational costs but also effectively detect the aforementioned plagiarism.

[0004] Text summarization offers an effective solution, achieving significant success by modeling document sentences, particularly in news and short articles. Generative and extractive methods are common in text summarization, and much current research is based on these two approaches. However, most current text summarization methods summarize short texts. When dealing with long texts, such as medical papers, they typically only process the abstract. This is because they still face challenges when handling long texts, such as the inability to extract more detailed structural features, maintain sentence coherence, and address grammatical issues. Therefore, simply summarizing the article from the abstract ignores much more detailed technical information and experimental details. Consequently, traditional models for summarizing short texts are not ideal for summarizing long texts and cannot comprehensively cover the important information conveyed in a given scientific text, thus failing to provide a comprehensive and effective comparison of the similarity of long texts. Summary of the Invention

[0005] The purpose of this invention is to provide a long text summarization method based on graph neural networks, which aims to solve the technical problem that existing pre-trained language models cannot capture the contextual relationships of long text sequences, thus failing to comprehensively construct inter-sentence relationships and ultimately failing to perform comprehensive and effective comparison of the similarity of long texts.

[0006] To achieve the above objectives, this invention provides a long text summarization method based on graph neural networks, comprising the following steps:

[0007] Obtain the source text and the corresponding summary tags;

[0008] The source text is preprocessed and then input into a pre-trained BERT to obtain the embedding representation of each sentence in the source text;

[0009] Each sentence of the source text is embedded into an inflated gated convolutional network and aggregated using a residual structure to obtain a new embedding representation of each sentence in the source text.

[0010] Construct a heterogeneous graph by linking each sentence of the source text with the words belonging to that sentence, construct an inter-sentence graph by semantic similarity, and iterate through a graph attention layer;

[0011] Each sentence of the source text is used to construct a heterogeneous graph with related topics. Then, a natural relationship graph between sentences is constructed through a topic model, and the graph attention layer is used for iteration.

[0012] The two different feature representations of each sentence in the source text are aggregated and input into the fully connected layer of the activation function to calculate the probability value of each sentence, obtain the predicted summary, and construct the summary loss function and the topic loss function.

[0013] The model loss function is calculated based on the summary loss function and the topic loss function;

[0014] The text summarization model is obtained by adjusting the weights of the entire neural network based on the model loss function.

[0015] The process of preprocessing the source text includes the following steps:

[0016] Insert start symbol [CLS] and end symbol [SEP] into all sentences in the source text, and represent each word in the sentence as a word vector;

[0017] Each input sequence, with a length not exceeding 512, is fed into the pre-trained BERT, and the sentence features are represented by the first vector feature of each sentence.

[0018] Specifically, the process of embedding each sentence of the source text into the dilated gated convolutional network involves sequentially inputting the feature representation of each sentence in the source text into the gated convolution to reduce the risk of gradient vanishing.

[0019] The output of the gated convolution is fed into the dilated convolutional network to learn longer contextual features. Finally, the features of each sentence are aggregated with the features passed through the dilated convolutional network in the form of a residual structure to obtain a new sentence feature representation.

[0020] The process of constructing a heterogeneous graph by associating each sentence of the source text with related topics includes the following steps:

[0021] Construct a heterogeneous graph that connects each sentence of the source text with the words belonging to that sentence, excluding conjunctions, interjections, and meaningless words;

[0022] Construct a graph of sentences from each sentence in the source text using BM25 semantic similarity.

[0023] The multi-level heterogeneous semantic similarity representation of each sentence is input into the graph attention network layer for iterative computation to generate a new representation.

[0024] The process of constructing a heterogeneous graph by associating each sentence of the source text with the words belonging to that sentence includes the following steps:

[0025] The source text is input into a neural topic model to obtain a set number of topics, and sentences in the source text are used to construct a heterogeneous graph with related topics;

[0026] Construct a natural relationship graph between sentences in the source text, categorizing them according to similar themes;

[0027] The multi-level natural relation representation of each sentence is input into the graph attention network layer for iterative computation to generate a new representation.

[0028] The process of aggregating the two different feature representations of each sentence in the source text, inputting them into a fully connected activation layer, calculating the probability value of each sentence to obtain a predicted summary, and constructing summary loss and topic loss functions includes the following steps:

[0029] The sentence with the highest probability value is selected as the predicted summary sentence to obtain the predicted summary;

[0030] A cross-entropy loss function is constructed based on the predicted source text summary and the tag summary;

[0031] A KL divergence topic loss function is constructed based on the word topics and document topics of the source text.

[0032] The summarization loss function is specifically as follows:

[0033] , Indicates sentence label, This indicates the predicted sentence label.

[0034] Specifically, the topic model loss function is as follows:

[0035] , It is the probability of topic z. It is the probability of topic z under word x. It is the probability of word x under topic z.

[0036] The model loss function is specifically as follows: +(1- ) ,in The summarization loss is the cross-entropy loss function. For the topic model loss function, This indicates the weight values ​​for adjusting the two loss functions mentioned above.

[0037] This invention provides a long text summarization method based on graph neural networks. The method involves preprocessing the source text to obtain word embeddings for each sentence, which are then input into a pre-trained BERT algorithm. The output of the BERT algorithm is then fed into a dilated gated convolutional network (DGCN). The BERT output and the DGCN output are aggregated in a residual structure to prevent gradient vanishing. Multi-level semantic similarity maps and multi-level natural relation maps are then constructed and passed through graph attention layers. After aggregation, sentence labels are predicted using activation functions and fully connected layers, and a summarization loss function is constructed. Finally, a model loss function is calculated based on the summarization loss function and the topic model loss function, enabling the extracted text summary to more comprehensively retain the important information of the source text. Attached Figure Description

[0038] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0039] Figure 1 This is a flowchart illustrating a long text summarization method based on graph neural networks according to the present invention.

[0040] Figure 2 This is a schematic diagram of the overall structure of the text summarization model of the present invention.

[0041] Figure 3 This is a schematic diagram of the pre-trained BERT model structure of the present invention.

[0042] Figure 4This is a schematic diagram of the dilatation-gated convolution model structure of the present invention.

[0043] Figure 5 This is the multi-level inter-sentence semantic similarity graph of the present invention.

[0044] Figure 6 This is the multi-level inter-sentence natural relationship diagram of the present invention. Detailed Implementation

[0045] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0046] Please see Figure 1 This invention provides a long text summarization method based on graph neural networks, comprising the following steps:

[0047] S1: Obtain the source text and the corresponding summary tags;

[0048] S2: Preprocess the source text and input it into the pre-trained BERT to obtain the embedding representation of each sentence in the source text;

[0049] S3: The embedding of each sentence in the source text is fed into an dilated gated convolutional network and aggregated in the form of a residual structure to obtain a new embedding representation of each sentence in the source text;

[0050] S4: Construct a heterogeneous graph by connecting each sentence of the source text with the words belonging to that sentence, construct an inter-sentence graph by semantic similarity, and iterate through a graph attention layer;

[0051] S5: Construct a heterogeneous graph for each sentence of the source text and related topics, then construct a natural relationship graph between sentences through a topic model, and iterate through a graph attention layer;

[0052] S6: Aggregate the two different feature representations of each sentence in the source text, input them into the fully connected layer of the activation function, calculate the probability value of each sentence, obtain the predicted summary, and construct the summary loss function and the topic loss function;

[0053] S7: The model loss function is calculated based on the summary loss function and the topic loss function;

[0054] S8: Adjust the weights of the entire neural network based on the model loss function to obtain the text summarization model.

[0055] The present invention will be further described below with reference to specific implementation steps:

[0056] S1: Obtain the source text and the corresponding summary tags;

[0057] In this context, source text refers to the data from which the text summarization model was trained. Summary labels refer to pre-defined labels used to identify the summary.

[0058] S2: Preprocess the source text and input it into the pre-trained BERT to obtain the embedding representation of each sentence in the source text;

[0059] The preprocessing involves inserting a start symbol [CLS] and an end symbol [SEP] into each sentence of the source text, and mapping each word in the sentence to a word vector, including the start and end symbols. The BERT output represents the corresponding sentence using the start symbol [CLS] of each sentence.

[0060] S3: The embedding of each sentence in the source text is fed into an dilated gated convolutional network and aggregated in the form of a residual structure to obtain a new embedding representation of each sentence in the source text.

[0061] The dilated gated convolutional network is a combination of dilated convolutional networks and gated convolutional networks. Due to the limitation of the length of the input sequence of the pre-trained language model, the dilated convolutional network can capture the contextual information that is truncated into unrelated parts due to the length exceeding 512, which can effectively alleviate the limitation of the fixed length of the input sequence. The gated convolutional network can improve performance, and the use of residual structure can prevent gradient vanishing.

[0062] S4: Construct a heterogeneous graph for each sentence in the source text and the words belonging to that sentence, construct an inter-sentence graph through semantic similarity, and iterate through a graph attention layer.

[0063] Each sentence and its constituent words form a heterogeneous graph, excluding conjunctions, interjections, and other words without practical meaning. After constructing the heterogeneous graph, a semantic similarity graph between sentences is built using BM25. First, graph attention between sentences and words is calculated and iterated several times. This iteration involves iterating between word-to-sentence and sentence-to-word graph attention. Then, graph attention between sentences is calculated, and the sentence-to-sentence iteration is repeated several times to capture the similarity relationships between sentences.

[0064] S5: Construct a heterogeneous graph for each sentence of the source text and related topics, then construct a natural relationship graph between sentences through a topic model, and iterate through a graph attention layer.

[0065] The topic model is constructed as follows: The source text is processed in a bag-of-words manner through two different fully connected networks to obtain the mean and covariance. A Gaussian distribution is then constructed using the bag-of-words approach, followed by an activation function to obtain topic probabilities and extract a certain number of topics. Next, a heterogeneous graph between sentences and topics is constructed using conditional probabilities, and a natural relationship graph between sentences is constructed using the topic conditional probabilities between sentences, effectively capturing synonym relationships. Then, graph attention between sentences and topics is calculated and iterated several times. This graph attention refers to the iterative graph attention from topic to sentence and from sentence to topic. Finally, graph attention between sentences is calculated to capture natural relationships between sentences.

[0066] S6: Aggregate the two different feature representations of each sentence in the source text, input them into the fully connected layer of the activation function, calculate the probability value of each sentence, obtain the predicted summary, and construct the summary loss and topic loss functions;

[0067] The probability value is calculated as the probability of selecting a sentence as the summary.

[0068] Specifically, the summarization loss function is:

[0069] , Indicates sentence label, The predicted sentence label is represented by the similarity calculated between the summary label (i.e., the theoretical summary) and the predicted summary (i.e., the summary actually output by the fully connected neural network).

[0070] Construct the loss function.

[0071] Specifically, the topic loss function is:

[0072] , It is the probability of topic z. Let be the probability of topic z given word x, and represent the encoder network of the topic model. Let be the probability of word x under topic z, and let represent the decoder network of the topic model.

[0073] S7: The model loss function is calculated based on the summary loss and topic loss functions.

[0074] The specific model loss function is as follows: +(1- ) ,in The summarization loss is the cross-entropy loss function. For the topic model loss function, To adjust the weight parameters of the two loss functions mentioned above.

[0075] S8: Adjust the weights of the entire neural network based on the model loss function to obtain the text summarization model.

[0076] in Figure 2 This refers to the overall structure of the model.

[0077] Furthermore, such as Figure 3 As shown, the source text is preprocessed and then input into a pre-trained BERT to obtain the embedding representation of each sentence in the source text. The specific steps are as follows:

[0078] In the source text, a start symbol [CLS] and an end symbol [SEP] are inserted at the beginning and end of each sentence, respectively. Then, each word in the sentence is represented by a pre-trained word vector, including the start and end symbols.

[0079] All word vectors are input into BERT, specifically as follows:

[0080]

[0081] in This represents the j-th word of the i-th sentence. and Let [CLS] and [SEP] represent the start symbol [CLS] and end symbol [SEP] of the i-th sentence, respectively. This indicates the hidden state of the corresponding symbol. This represents the i-th sentence.

[0082] Furthermore, such as Figure 4 As shown, each sentence of the source text is embedded into the dilated gated convolutional network, and the specific steps are as follows:

[0083] First, it passes through a gated convolutional neural network using a residual structure. The specific formula is as follows:

[0084]

[0085] in for The activation function, where H is the sentence representation output by BERT. Conv1D1 and Conv1D2 convolutional layers use the same window and number of kernels, but the weights of the two convolutional layers are not shared.

[0086] The output Y of the gated convolution is input into a dilated convolution, where the dilation rate of each layer is set to 1, 2, and 4. The window size is set to 3 for all layers to obtain new features. .

[0087] Furthermore, such as Figure 5As shown, a heterogeneous graph is constructed by linking each sentence of the source text with the words belonging to that sentence, and an inter-sentence graph is constructed using semantic similarity. The graph is then iterated through a graph attention layer. The specific steps are as follows:

[0088] The heteromorphism diagram of sentences and words belonging to those sentences does not include pause words, conjunctions, or other words without actual meaning.

[0089] The semantic similarity between sentences is constructed using the classic BM25 algorithm. First, graph attention between sentences and words is calculated, and then iterated a certain number of times. This iteration refers to graph attention calculations from words to sentences and from sentences to words. Finally, graph attention between sentences is calculated.

[0090] The word vectors use pre-trained vectors, and the entire vocabulary is contained in glove.42B.300d.txt. Meaningless words are filtered out. The sentence vectors are the outputs of BERT and DGCNN for each sentence's start symbol.

[0091] The formula for calculating graph attention from words to sentences is as follows:

[0092]

[0093] The formula for calculating graph attention from sentence to word is as follows:

[0094]

[0095] in and These refer to sentence and word vectors, respectively. Indicates the connection between the two. , , , For trainable parameters, For activation function, Calculate graph attention from words to sentences. Calculate the graph attention from sentence to word.

[0096]

[0097]

[0098] It is a word to sentence The weights, where Representation and sentence vector A set of connected words, It is a sentence To words The weight, Representation and word vectors A collection of connected sentences.

[0099]

[0100]

[0101] This invention employs multi-head attention, where the number is K, and || denotes aggregation. For activation function, , For trainable parameters Indicates and sentences A set of connected word vectors. With words A set of connected sentences. After multi-head attention computation, new sentence matrices S and word matrices Z are generated.

[0102] To prevent gradient vanishing, a residual structure was added:

[0103]

[0104] The initial iteration from word to sentence is:

[0105]

[0106] In the initial iteration from word to sentence, Let S be the embedding matrix of the sentence, and let S be the matrix. The query matrix, key matrix, and value matrix are represented by word embedding matrices. express.

[0107] The (t+1)th iteration from word to sentence:

[0108]

[0109] To prevent gradient vanishing, a residual structure was added:

[0110]

[0111] Sentence to word iteration t+1

[0112]

[0113] To prevent gradient vanishing, a residual structure was added:

[0114]

[0115] We obtain new feature representations of the sentences and then perform graph attention computation between the sentences:

[0116]

[0117] Where Y is the set of sentences connected to sentence s through semantic similarity.

[0118] To prevent gradient vanishing, a residual structure was added:

[0119]

[0120] in, To distinguish new features of natural sentences, we use [the following] here. This represents the new feature of the final iteration.

[0121] Furthermore, such as Figure 6 As shown, a heterogeneous graph is constructed by linking each sentence of the source text with related topics, and an inter-sentence graph is constructed using conditional topic probabilities. The graph is then iterated through a graph attention layer. The specific steps are as follows:

[0122] In this process, the topic model is constructed by taking the source text as input to obtain the bag-of-words. The input is a bag-of-words word that passes through two different linear fully connected layers with activation functions to obtain the mean. and standard deviation The details are as follows:

[0123]

[0124]

[0125] in Activation function , It is a fully connected layer.

[0126] Construct a topic Gaussian distribution using the mean and variance. Then obtain the topic distribution. ,

[0127] ,in It is a topic-word distribution matrix, where To represent the relevance between word i and topic j, through... Establish relationships between topics and words.

[0128] Specifically, obtain a certain number of topics and related weights :

[0129]

[0130]

[0131] in It is a The linear transition function of the activation function, for transpose, Let be the probability of the i-th topic. It is the i-th topic.

[0132] First, the sentence is connected to its related topic. Then, graph attention between the sentence and the topic is calculated and iterated a certain number of times. The iteration refers to the graph attention from topic to sentence and the graph attention from sentence to topic. New embedding representations U and T of the sentence and topic are obtained. Further, graph attention between sentences is calculated.

[0133] The formula for calculating the attention between sentences and topics is as follows:

[0134]

[0135] The formula for calculating the attention between topics and sentences is as follows:

[0136]

[0137] in and Representing sentence and topic vectors, It is the connection between them. , , , For trainable parameters, For activation function, Calculate the graph attention between sentences and topics. Calculate the graph attention between topics and sentences.

[0138]

[0139]

[0140] yes The weights, where Representation and sentence vector A collection of interconnected themes. yes The weight, It is related to topic vectors A set of connected sentences.

[0141]

[0142]

[0143] This invention employs multi-head attention, where the number is K, and || denotes aggregation. For activation function, For trainable parameters It is related to sentence vectors A set of connected topic vectors. It is a topic vector A set of connected sentence vectors.

[0144] To prevent gradient vanishing, a residual structure method was also employed.

[0145]

[0146] The initial iteration from topic to sentence is:

[0147]

[0148] In the initial iteration, Let U be the embedding matrix of the sentences, and let U be the query matrix. for , representing the key matrix and the value matrix.

[0149] The (t+1)th iteration from topic to sentence:

[0150]

[0151] To prevent gradient vanishing, a residual structure was added:

[0152]

[0153] Iteration from sentence to topic (t+1th word)

[0154]

[0155] To prevent gradient vanishing, a residual structure was added:

[0156]

[0157] We obtain new feature representations of sentences and perform sentence-to-sentence graph attention computation:

[0158]

[0159] Where N is the set of sentences connected to sentence s through a topic, and the specific formula for calculating the connection between two sentences is as follows: The natural relationship between two sentences is established by using the KL divergence value.

[0160] To prevent gradient vanishing, a residual structure was added:

[0161]

[0162] in, This is a new feature for the corresponding sentence; to distinguish it from similarity features, we use [the following]. Features representing the last iteration .

[0163] Furthermore, the two different feature representations of each sentence in the source text are aggregated and input into the fully connected layer of the activation function to calculate the probability value of each sentence, obtain the predicted summary, and construct the summary loss and topic loss functions;

[0164] Among them, new features and according to A certain number of sentences are selected as text summaries based on their probability.

[0165] It is a linear fully connected layer. For activation function, This represents the probability value of the corresponding sentence sent.

[0166] The summarization loss function is:

[0167] , Indicates sentence label, The predicted sentence label is represented by the similarity calculated between the summary label (i.e., the theoretical summary) and the predicted summary (i.e., the summary actually output by the fully connected neural network).

[0168] To construct the loss function.

[0169] The topic loss function is:

[0170]

[0171] in, It is the probability of topic z. This represents the probability value of topic x. This represents the probability value of a word under topic z.

[0172] Furthermore, the model loss function is calculated based on the summary loss and topic loss functions;

[0173] The specific model loss function is as follows: +(1- ) ,in The summarization loss is the cross-entropy loss function. For the topic model loss function, To adjust the weight parameters of the two loss functions mentioned above.

[0174] Furthermore, the weights of the entire neural network are adjusted based on the model loss function to obtain a long text summarization model.

[0175] The above description discloses only one preferred embodiment of the present invention, and should not be construed as limiting the scope of the present invention. Those skilled in the art will understand that all or part of the processes of the above embodiments can be implemented, and equivalent changes made in accordance with the claims of the present invention are still within the scope of the invention.

Claims

1. A method for generating long text summaries based on graph neural networks, characterized in that, Includes the following steps: Obtain the source text and the corresponding summary tags; The source text is preprocessed and then input into a pre-trained BERT to obtain the embedding representation of each sentence in the source text; Each sentence of the source text is embedded into an inflated gated convolutional network and aggregated using a residual structure to obtain a new embedding representation of each sentence in the source text. Dilated convolutional networks are used to capture contextual information that is truncated into unrelated parts due to lengths exceeding 512, thus alleviating the limitation of fixed-length input order. Construct a heterogeneous graph by linking each sentence of the source text with the words belonging to that sentence, construct an inter-sentence graph by semantic similarity, and iterate through a graph attention layer; The process of constructing a heterogeneous graph of each sentence in the source text with the words belonging to that sentence includes the following steps: Construct a heterogeneous graph that connects each sentence of the source text with the words belonging to that sentence, excluding conjunctions, interjections, and meaningless words; A graph of sentences in the source text is constructed using BM25 semantic similarity. The multi-level heterogeneous semantic similarity representation of each sentence is input into the graph attention network layer for iterative computation to obtain a new representation. Each sentence of the source text is used to construct a heterogeneous graph with related topics. Then, a natural relationship graph between sentences is constructed through a topic model, and the graph attention layer is used for iteration. The process of constructing a heterogeneous graph of each sentence in the source text with related topics includes the following steps: The source text is input into a neural topic model to obtain a set number of topics, and sentences in the source text are used to construct a heterogeneous graph with related topics; Construct a natural relationship graph between sentences in the source text, categorizing each sentence according to similar themes; The multi-level natural relation representation of each sentence is input into the graph attention network layer for iterative computation to obtain a new representation; The semantic similarity feature representation and natural relation graph feature representation of each sentence in the source text are aggregated and input into a fully connected layer of the activation function. The probability value of each sentence is calculated to obtain a predicted summary. The summary loss function and topic loss function are then constructed, including the following steps: The sentence with the highest probability value is selected as the predicted summary sentence to obtain the predicted summary; A cross-entropy loss function is constructed based on the predicted source text summary and the tag summary; Construct a KL divergence topic loss function based on the word topics and document topics of the source text; The model loss function is calculated based on the summary loss function and the topic loss function; The text summarization model is obtained by adjusting the weights of the entire neural network based on the model loss function.

2. The long text summarization method based on graph neural networks as described in claim 1, characterized in that, The process of preprocessing the source text includes the following steps: Insert start symbol [CLS] and end symbol [SEP] into all sentences in the source text, and represent each word in the sentence as a word vector; Each input sequence, with a length not exceeding 512, is fed into the pre-trained BERT, and the sentence features are represented by the first vector feature of each sentence.

3. The long text summarization method based on graph neural networks as described in claim 1, characterized in that, The process of embedding each sentence of the source text into the dilated gated convolutional network specifically involves inputting the feature representation of each sentence in the source text into the gated convolution in sequence to reduce the risk of gradient vanishing. The output of the gated convolution is fed into the dilated convolutional network to learn longer contextual features. Finally, the features of each sentence are aggregated with the features passed through the dilated convolutional network in the form of a residual structure to obtain a new sentence feature representation.

4. The long text summarization method based on graph neural networks as described in claim 1, characterized in that, The specific summary loss function is as follows: , Indicates sentence label, This indicates the predicted sentence label.

5. The long text summarization method based on graph neural networks as described in claim 4, characterized in that, The topic model loss function is specifically as follows: , It is the probability of topic z. It is the probability of topic z under word x. It is the probability of word x under topic z.

6. The long text summarization method based on graph neural networks as described in claim 5, characterized in that, The model loss function is specifically as follows: +(1- ) ,in The summarization loss is the cross-entropy loss function. For the topic model loss function, This indicates the weight values ​​for adjusting the two loss functions mentioned above.

Images